Rory Sayres

Rory Sayres, Kejia Chen, Ayush Jain et al.

Patient-managed Personal Health Records (PHRs) promises to empower patients to better understand their health; but information in the record is complex, potentially hindering insights. In this study, we assess the potential of large language models (LLMs, Gemini 3.0 Flash) to provide helpful answers to user health queries, when provided clinical data from PHRs as context. A total of 2,257 user queries were drawn from 3 different distributions to represent patient questions: shorter web search queries, longer questions derived from templates of chatbot conversations, and questions patients asked to their healthcare team (patient calls). Queries were matched with de-identified PHRs (from a pool of 1,945). Gemini responses were generated (1) without PHR context; (2) with a basic summary of demographics, conditions, and medications; (3) with full, extensive clinical notes. For evaluation, we leveraged an existing rating framework (SHARP), and developed a new framework for specific error modes when interpreting PHRs. Evaluation was performed using autoraters for the full set, and with clinician ratings for a subset (n=95), with both sets of raters knowing the full PHR context. We see significant improvements in the helpfulness of answers to all question types with PHR data (p < 0.001, paired t-test). We also observe potential gains in safety, accuracy, relevance and personalization of answers. Our PHR evaluation framework further identifies gaps in LLM understanding of particular aspects of complex PHRs, such as temporal disorientation, and rare but meaningful confabulations. These results suggest potential for PHR data to help people with a wide range of user needs; and provide a framework for monitoring for gaps in LLM answers based on PHR context. This study motivates further work to assess and realize potential benefits to users from understanding their health records.

Stephen R. Pfohl, Heather Cole-Lewis, Rory Sayres et al.

Large language models (LLMs) hold promise to serve complex health information needs but also have the potential to introduce harm and exacerbate health disparities. Reliably evaluating equity-related model failures is a critical step toward developing systems that promote health equity. We present resources and methodologies for surfacing biases with potential to precipitate equity-related harms in long-form, LLM-generated answers to medical questions and conduct a large-scale empirical case study with the Med-PaLM 2 LLM. Our contributions include a multifactorial framework for human assessment of LLM-generated answers for biases, and EquityMedQA, a collection of seven datasets enriched for adversarial queries. Both our human assessment framework and dataset design process are grounded in an iterative participatory approach and review of Med-PaLM 2 answers. Through our empirical study, we find that our approach surfaces biases that may be missed via narrower evaluation approaches. Our experience underscores the importance of using diverse assessment methodologies and involving raters of varying backgrounds and expertise. While our approach is not sufficient to holistically assess whether the deployment of an AI system promotes equitable health outcomes, we hope that it can be leveraged and built upon towards a shared goal of LLMs that promote accessible and equitable healthcare.

Rajeev V. Rikhye, Aaron Loh, Grace Eunhae Hong et al.

Recently, there has been great progress in the ability of artificial intelligence (AI) algorithms to classify dermatological conditions from clinical photographs. However, little is known about the robustness of these algorithms in real-world settings where several factors can lead to a loss of generalizability. Understanding and overcoming these limitations will permit the development of generalizable AI that can aid in the diagnosis of skin conditions across a variety of clinical settings. In this retrospective study, we demonstrate that differences in skin condition distribution, rather than in demographics or image capture mode are the main source of errors when an AI algorithm is evaluated on data from a previously unseen source. We demonstrate a series of steps to close this generalization gap, requiring progressively more information about the new source, ranging from the condition distribution to training data enriched for data less frequently seen during training. Our results also suggest comparable performance from end-to-end fine tuning versus fine tuning solely the classification layer on top of a frozen embedding model. Our approach can inform the adaptation of AI algorithms to new settings, based on the information and resources available.

Rory Sayres, Yuexing Hao, Abbi Ward et al. · deepmind

Navigating health questions can be daunting in the modern information landscape. Large language models (LLMs) may provide tailored, accessible information, but also risk being inaccurate, biased or misleading. We present insights from 4 mixed-methods studies (total N=163), examining how people interact with LLMs for their own health questions. Qualitative studies revealed the importance of context-seeking in conversational AIs to elicit specific details a person may not volunteer or know to share. Context-seeking by LLMs was valued by participants, even if it meant deferring an answer for several turns. Incorporating these insights, we developed a "Wayfinding AI" to proactively solicit context. In a randomized, blinded study, participants rated the Wayfinding AI as more helpful, relevant, and tailored to their concerns compared to a baseline AI. These results demonstrate the strong impact of proactive context-seeking on conversational dynamics, and suggest design patterns for conversational AI to help navigate health topics.

Karan Singhal, Tao Tu, Juraj Gottweis et al.

Recent artificial intelligence (AI) systems have reached milestones in "grand challenges" ranging from Go to protein-folding. The capability to retrieve medical knowledge, reason over it, and answer medical questions comparably to physicians has long been viewed as one such grand challenge. Large language models (LLMs) have catalyzed significant progress in medical question answering; Med-PaLM was the first model to exceed a "passing" score in US Medical Licensing Examination (USMLE) style questions with a score of 67.2% on the MedQA dataset. However, this and other prior work suggested significant room for improvement, especially when models' answers were compared to clinicians' answers. Here we present Med-PaLM 2, which bridges these gaps by leveraging a combination of base LLM improvements (PaLM 2), medical domain finetuning, and prompting strategies including a novel ensemble refinement approach. Med-PaLM 2 scored up to 86.5% on the MedQA dataset, improving upon Med-PaLM by over 19% and setting a new state-of-the-art. We also observed performance approaching or exceeding state-of-the-art across MedMCQA, PubMedQA, and MMLU clinical topics datasets. We performed detailed human evaluations on long-form questions along multiple axes relevant to clinical applications. In pairwise comparative ranking of 1066 consumer medical questions, physicians preferred Med-PaLM 2 answers to those produced by physicians on eight of nine axes pertaining to clinical utility (p < 0.001). We also observed significant improvements compared to Med-PaLM on every evaluation axis (p < 0.001) on newly introduced datasets of 240 long-form "adversarial" questions to probe LLM limitations. While further studies are necessary to validate the efficacy of these models in real-world settings, these results highlight rapid progress towards physician-level performance in medical question answering.

Alexander D'Amour, Katherine Heller, Dan Moldovan et al.

ML models often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification as a key reason for these failures. An ML pipeline is underspecified when it can return many predictors with equivalently strong held-out performance in the training domain. Underspecification is common in modern ML pipelines, such as those based on deep learning. Predictors returned by underspecified pipelines are often treated as equivalent based on their training domain performance, but we show here that such predictors can behave very differently in deployment domains. This ambiguity can lead to instability and poor model behavior in practice, and is a distinct failure mode from previously identified issues arising from structural mismatch between training and deployment domains. We show that this problem appears in a wide variety of practical ML pipelines, using examples from computer vision, medical imaging, natural language processing, clinical risk prediction based on electronic health records, and medical genomics. Our results show the need to explicitly account for underspecification in modeling pipelines that are intended for real-world deployment in any domain.

Joy Hsu, Sonia Phene, Akinori Mitani et al.

As machine learning has become increasingly applied to medical imaging data, noise in training labels has emerged as an important challenge. Variability in diagnosis of medical images is well established; in addition, variability in training and attention to task among medical labelers may exacerbate this issue. Methods for identifying and mitigating the impact of low quality labels have been studied, but are not well characterized in medical imaging tasks. For instance, Noisy Cross-Validation splits the training data into halves, and has been shown to identify low-quality labels in computer vision tasks; but it has not been applied to medical imaging tasks specifically. In this work we introduce Stratified Noisy Cross-Validation (SNCV), an extension of noisy cross validation. SNCV can provide estimates of confidence in model predictions by assigning a quality score to each example; stratify labels to handle class imbalance; and identify likely low-quality labels to analyze the causes. We assess performance of SNCV on diagnosis of glaucoma suspect risk from retinal fundus photographs, a clinically important yet nuanced labeling task. Using training data from a previously-published deep learning model, we compute a continuous quality score (QS) for each training example. We relabel 1,277 low-QS examples using a trained glaucoma specialist; the new labels agree with the SNCV prediction over the initial label >85% of the time, indicating that low-QS examples mostly reflect labeler errors. We then quantify the impact of training with only high-QS labels, showing that strong model performance may be obtained with many fewer examples. By applying the method to randomly sub-sampled training dataset, we show that our method can reduce labelling burden by approximately 50% while achieving model performance non-inferior to using the full dataset on multiple held-out test sets.

Arunachalam Narayanaswamy, Subhashini Venugopalan, Dale R. Webster et al.

Model explanation techniques play a critical role in understanding the source of a model's performance and making its decisions transparent. Here we investigate if explanation techniques can also be used as a mechanism for scientific discovery. We make three contributions: first, we propose a framework to convert predictions from explanation techniques to a mechanism of discovery. Second, we show how generative models in combination with black-box predictors can be used to generate hypotheses (without human priors) that can be critically examined. Third, with these techniques we study classification models for retinal images predicting Diabetic Macular Edema (DME), where recent work showed that a CNN trained on these images is likely learning novel features in the image. We demonstrate that the proposed framework is able to explain the underlying scientific mechanism, thus bridging the gap between the model's performance and human understanding.

Sonia Phene, R. Carter Dunn, Naama Hammel et al.

Glaucoma is the leading cause of preventable, irreversible blindness world-wide. The disease can remain asymptomatic until severe, and an estimated 50%-90% of people with glaucoma remain undiagnosed. Glaucoma screening is recommended for early detection and treatment. A cost-effective tool to detect glaucoma could expand screening access to a much larger patient population, but such a tool is currently unavailable. We trained a deep learning algorithm using a retrospective dataset of 86,618 images, assessed for glaucomatous optic nerve head features and referable glaucomatous optic neuropathy (GON). The algorithm was validated using 3 datasets. For referable GON, the algorithm had an AUC of 0.945 (95% CI, 0.929-0.960) in dataset A (1205 images, 1 image/patient; 18.1% referable), images adjudicated by panels of Glaucoma Specialists (GSs); 0.855 (95% CI, 0.841-0.870) in dataset B (9642 images, 1 image/patient; 9.2% referable), images from Atlanta Veterans Affairs Eye Clinic diabetic teleretinal screening program; and 0.881 (95% CI, 0.838-0.918) in dataset C (346 images, 1 image/patient; 81.7% referable), images from Dr. Shroff's Charity Eye Hospital's glaucoma clinic. The algorithm showed significantly higher sensitivity than 7 of 10 graders not involved in determining the reference standard, including 2 of 3 GSs, and showed higher specificity than 3 graders, while remaining comparable to others. For both GSs and the algorithm, the most crucial features related to referable GON were: presence of vertical cup-to-disc ratio of 0.7 or more, neuroretinal rim notching, retinal nerve fiber layer defect, and bared circumlinear vessels. An algorithm trained on fundus images alone can detect referable GON with higher sensitivity than and comparable specificity to eye care providers. The algorithm maintained good performance on an independent dataset with diagnoses based on a full glaucoma workup.

Paisan Raumviboonsuk, Jonathan Krause, Peranut Chotcomwongse et al.

Deep learning algorithms have been used to detect diabetic retinopathy (DR) with specialist-level accuracy. This study aims to validate one such algorithm on a large-scale clinical population, and compare the algorithm performance with that of human graders. 25,326 gradable retinal images of patients with diabetes from the community-based, nation-wide screening program of DR in Thailand were analyzed for DR severity and referable diabetic macular edema (DME). Grades adjudicated by a panel of international retinal specialists served as the reference standard. Across different severity levels of DR for determining referable disease, deep learning significantly reduced the false negative rate (by 23%) at the cost of slightly higher false positive rates (2%). Deep learning algorithms may serve as a valuable tool for DR screening.

Maithra Raghu, Katy Blumer, Rory Sayres et al.

The issue of disagreements amongst human experts is a ubiquitous one in both machine learning and medicine. In medicine, this often corresponds to doctor disagreements on a patient diagnosis. In this work, we show that machine learning models can be trained to give uncertainty scores to data instances that might result in high expert disagreements. In particular, they can identify patient cases that would benefit most from a medical second opinion. Our central methodological finding is that Direct Uncertainty Prediction (DUP), training a model to predict an uncertainty score directly from the raw patient features, works better than Uncertainty Via Classification, the two-step process of training a classifier and postprocessing the output distribution to give an uncertainty score. We show this both with a theoretical result, and on extensive evaluations on a large scale medical imaging application.

Been Kim, Martin Wattenberg, Justin Gilmer et al.

The interpretation of deep learning models is a challenge due to their size, complexity, and often opaque internal state. In addition, many systems, such as image classifiers, operate on low-level features rather than high-level concepts. To address these challenges, we introduce Concept Activation Vectors (CAVs), which provide an interpretation of a neural net's internal state in terms of human-friendly concepts. The key idea is to view the high-dimensional internal state of a neural net as an aid, not an obstacle. We show how to use CAVs as part of a technique, Testing with CAVs (TCAV), that uses directional derivatives to quantify the degree to which a user-defined concept is important to a classification result--for example, how sensitive a prediction of "zebra" is to the presence of stripes. Using the domain of image classification as a testing ground, we describe how CAVs may be used to explore hypotheses and generate insights for a standard image classification network as well as a medical application.